Skip to content

Blog

  • in AI
  • 3 min read

Escaping ouroboros: build systems that maximize interaction

The early adopters of the AI wave had a gift they didn't fully appreciate: clean ground truth. The pre-2018 web was messy but human. The early quants traded against human intuition. But that world is fading.

We have entered the era of the data ouroboros: models trained on the outputs of models. If you keep training on the raw fire hose, you aren't learning reality. You're photocopying a photocopy.

The way out is not more data. It's better interaction.

This piece explains why the convergence of quant shops and AI labs matters, and how to build systems that maximize interaction to escape the loop.

  • in AI
  • 4 min read

Backpressure: making subagents 10x more effective

Subagents are great at exploring, but terrible at stopping. If you've ever spun up a helper agent and watched it spray dozens of tool calls, you know the pain: high latency, noisy output, and a growing chance it drifts into nonsense.

The fix isn't more prompts. It's backpressure.

In this article, I'll walk through how I made subagents roughly 10x more effective by applying three levers:

  • Reduce tool calls without hurting accuracy
  • Enrich error logs so agents recover faster
  • Minimize total steps before they go runaway

With reference to projects I've worked on, you'll get to learn what it takes to evolve a subagent out-of-the-box to to a successful, token-efficient one.

  • in AI, Voice
  • 6 min read

How to perform LLM Evals

Evaluating LLM pipelines/workflows improvements is key to improving our AI systems. With limited time and resources, oftentimes the blocker is: overthinking them. In this article, I'll talk through a couple simple evals for benchmarking your improvements, based on work I've done previously.

The Curse of Overthinking

I've learnt that evals really can be as simple as an assert statement. The goal here is doing quick "smoke tests" to ensure that your pipeline is working as expected, whilst accounting for stochasticity. From that point, complexity is incrementally earned by structuring your evals around your most common & important failure modes.

If these failure modes aren't immediately apparent to you yet, then hunt for them first.

  • in AI, Voice
  • 4 min read

Working with Voice AI is Easy, Try This

Getting started with Voice AI can be easy. It's important to start simple to progressively build an understanding of what happens under the hood. By doing so, we can build a solid foundation for more complex applications. In addition, starting simple and adding complexity slowly helps us compare and appreciate the delta between demo-land & deployments into production.

Let's explore the following:

  1. Simple Speech-to-Text (STT)
  2. STT + Structured Outputs (e.g. JSON, Pydantic)
  3. Evals for Audio

For starters, let's keep it simple with a basic speech-to-text provider.

  • in Personal
  • 4 min read

My Wild Ride On A Floating Factory

Growing up in Singapore, I knew nothing about offshore oil and gas. My dad used to work in maritime industries, and I'd zone out every time he talked about shipping and charts. Little did I know I'd one day find myself on a massive floating vessel that's basically a city in the middle of the ocean.

The one I went on recently was the FPSO One Guyana, the largest of its kind globally. As you can see here, even the word 'massive' doesn't do it justice when you compare it's size to the smaller boats beside it.

One Guyana FPSO